Reinforcement learning continuously controls quantum error correction on Google’s Willow processor
Quantum error correction is the central challenge standing between today’s noisy quantum processors and the fault-tolerant quantum computers needed for practical applications. Correcting errors is one thing; keeping the correction system stable as the hardware drifts over time, a near-inevitable consequence of real physical devices, is another.
A collaboration between Google Quantum AI and Google DeepMind, published July 8 in Nature, has now demonstrated that reinforcement learning can solve the latter problem. On a Willow superconducting processor with 105 qubits, an RL agent continuously steered the control parameters of the error-correction system, achieving a 3.5-fold improvement in logical stability against drift and pushing surface code logical error rates to a new record: 7.72 × 10⁻⁴ per cycle.
The problem with drift
Quantum error correction (QEC) works by encoding a single logical qubit into many physical qubits and repeatedly measuring error-detecting “stabilizer” circuits. On the surface code, currently the most widely studied QEC scheme, the standard approach is to calibrate the control parameters (gate amplitudes, frequencies, coupling strengths) once, then run the computation.
But physical hardware drifts. Temperature fluctuations, two-level-system defects shifting in the dielectric, and other environmental changes cause control parameters to wander over time. Traditional recalibration requires stopping computation, which is unacceptable for algorithms that may run for days or months.
How RL solves this
The key insight: the same stabilizer measurements used to detect and correct qubit errors can be repurposed as a learning signal for an RL agent. The surrogate objective is the average rate of error-detection events across all detectors, which is proportional, through a known factor depending on code distance, to the logical error rate itself.
The agent uses a parameter-exploring policy gradient (PGPE) algorithm with a multivariate Gaussian policy. Because the factor-graph structure of the surface code means each detector depends only on local control parameters within its “detecting region,” gradient updates are naturally sparse and efficient, demonstrated in simulations up to distance-15 surface code with approximately 40,000 control parameters.
The system also exploits entropy regularization: in non-stationary (drift) settings, keeping the policy distribution from collapsing enables perpetual adaptation, rather than converging to a fixed set of parameters.
The results
On the Willow processor, the RL agent managed more than 1,000 control parameters simultaneously. The results under injected artificial drift (step, sinusoidal, and stroboscopic perturbations to CZ coupling, XY amplitude, and frequency):
- 2.4× improvement in logical error rate stability (standard deviation) with controller steering alone
- 3.5× improvement when also steering the decoder (reweighting the minimum-weight perfect matching graph)
- 31% reduction in mean logical error rate with combined steering
- Characteristic response time: approximately 130 epochs to recover from a step drift
Under natural drift, the uncontrolled, real-world variation the processor experiences during normal operation, the system achieved approximately 4 dB suppression of low-frequency logical error rate fluctuations.
The absolute logical error rates achieved represent new records for their respective code types:
- Distance-7 surface code: 7.72 × 10⁻⁴ per cycle (with the AlphaQubit2 neural decoder)
- Distance-5 colour code: 8.19 × 10⁻³ per cycle (with the Tesseract most-likely-error decoder)
The previous surface code record, set by the same Willow processor in December 2024, was 0.143% per cycle. This work improves that by roughly a factor of two.
A new paradigm for calibration
Perhaps the most striking demonstration: the RL agent could recover the full performance of the error-correction system starting from randomized initial control parameters, parameters essentially chosen at random. This suggests RL could eventually replace the traditional calibration stack entirely, not just augment it.
The authors frame the work as unifying the traditionally separate processes of calibration and computation. Rather than calibrate, then compute, then recalibrate, the system learns continuously, adapting to hardware changes in real time without ever interrupting the computation.
Sources:
1. Google Quantum AI and Collaborators. “Reinforcement learning control of quantum error correction.” Nature (2026). DOI: 10.1038/s41586-026-10759-2
2. Also on arXiv: 2511.08493 [quant-ph]
3. Data available on Zenodo: 10.5281/zenodo.17566521